LEAKY-INTEGRATOR RECONSTRUCTION: TAMING ERROR ACCUMULATION IN RECURSIVE DIFFERENCED TIME-SERIES FORECASTING

Battle Count: 0

Highly relevant for long-horizon financial forecasting. Recursive differencing is common in handling non-stationary financial data (prices/returns). This method significantly reduces error accumulation in long-term predictions (up to 51% improvement at H=336), which is critical for strategic asset allocation and long-term risk modeling.

Extremely low complexity. It requires no retraining, no new parameters, and is implemented as a simple exponential moving average (first-order IIR filter) on the predicted increments. It is described as a 'two-line change'.

The method is a simple two-line code change (applying a fixed gamma=0.9 filter) to existing models. The paper provides extensive empirical results across 20 datasets and 9 architectures, with clear mathematical definitions and synthetic validation.